gsdx sw: add extrathreads_height to control the quantity of pixels processed by a thread

Value could range from 1 to 9. Default is 4 and it is potentially the
best option.  Feel free to test some values on your system, behavior
might depends on the core number and thread number

Value is exponential so 4 is 2 times more pixels than 3.

Small value increased thread overhead, big value increase wait/sync latency
This commit is contained in:
Gregory Hainaut 2016-07-10 17:44:56 +02:00
parent 91eccb7bab
commit e642bbc426
3 changed files with 49 additions and 35 deletions

View File

@ -24,14 +24,21 @@
#include "stdafx.h"
#include "GSRasterizer.h"
// - for more threads screen segments should be smaller to better distribute the pixels
// - but not too small to keep the threading overhead low
// - ideal value between 3 and 5, or log2(64 / number of threads)
#define THREAD_HEIGHT 4
int GSRasterizerData::s_counter = 0;
static int compute_best_thread_height(int threads) {
// - for more threads screen segments should be smaller to better distribute the pixels
// - but not too small to keep the threading overhead low
// - ideal value between 3 and 5, or log2(64 / number of threads)
int th = theApp.GetConfigI("extrathreads_height");
if (th > 0 && th < 9)
return th;
else
return 4;
}
GSRasterizer::GSRasterizer(IDrawScanline* ds, int id, int threads, GSPerfMon* perfmon)
: m_perfmon(perfmon)
, m_ds(ds)
@ -40,14 +47,16 @@ GSRasterizer::GSRasterizer(IDrawScanline* ds, int id, int threads, GSPerfMon* pe
{
memset(&m_pixels, 0, sizeof(m_pixels));
m_thread_height = compute_best_thread_height(threads);
m_edge.buff = (GSVertexSW*)vmalloc(sizeof(GSVertexSW) * 2048, false);
m_edge.count = 0;
m_scanline = (uint8*)_aligned_malloc((2048 >> THREAD_HEIGHT) + 16, 64);
m_scanline = (uint8*)_aligned_malloc((2048 >> m_thread_height) + 16, 64);
int row = 0;
while(row < (2048 >> THREAD_HEIGHT))
while(row < (2048 >> m_thread_height))
{
for(int i = 0; i < threads; i++, row++)
{
@ -69,15 +78,15 @@ bool GSRasterizer::IsOneOfMyScanlines(int top) const
{
ASSERT(top >= 0 && top < 2048);
return m_scanline[top >> THREAD_HEIGHT] != 0;
return m_scanline[top >> m_thread_height] != 0;
}
bool GSRasterizer::IsOneOfMyScanlines(int top, int bottom) const
{
ASSERT(top >= 0 && top < 2048 && bottom >= 0 && bottom < 2048);
top = top >> THREAD_HEIGHT;
bottom = (bottom + (1 << THREAD_HEIGHT) - 1) >> THREAD_HEIGHT;
top = top >> m_thread_height;
bottom = (bottom + (1 << m_thread_height) - 1) >> m_thread_height;
while(top < bottom)
{
@ -92,13 +101,13 @@ bool GSRasterizer::IsOneOfMyScanlines(int top, int bottom) const
int GSRasterizer::FindMyNextScanline(int top) const
{
int i = top >> THREAD_HEIGHT;
int i = top >> m_thread_height;
if(m_scanline[i] == 0)
{
while(m_scanline[++i] == 0);
top = i << THREAD_HEIGHT;
top = i << m_thread_height;
}
return top;
@ -109,10 +118,10 @@ void GSRasterizer::Queue(const shared_ptr<GSRasterizerData>& data)
Draw(data.get());
}
int GSRasterizer::GetPixels(bool reset)
int GSRasterizer::GetPixels(bool reset)
{
int pixels = m_pixels.sum;
if(reset)
{
m_pixels.sum = 0;
@ -136,7 +145,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
const GSVertexSW* vertex = data->vertex;
const GSVertexSW* vertex_end = data->vertex + data->vertex_count;
const uint32* index = data->index;
const uint32* index_end = data->index + data->index_count;
@ -156,7 +165,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
{
DrawPoint<true>(vertex, data->vertex_count, index, data->index_count);
}
else
else
{
DrawPoint<false>(vertex, data->vertex_count, index, data->index_count);
}
@ -164,7 +173,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
break;
case GS_LINE_CLASS:
if(index != NULL)
{
do {DrawLine(vertex, index); index += 2;}
@ -179,7 +188,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
break;
case GS_TRIANGLE_CLASS:
if(index != NULL)
{
do {DrawTriangle(vertex, index); index += 3;}
@ -194,7 +203,7 @@ void GSRasterizer::Draw(GSRasterizerData* data)
break;
case GS_SPRITE_CLASS:
if(index != NULL)
{
do {DrawSprite(vertex, index); index += 2;}
@ -274,7 +283,7 @@ void GSRasterizer::DrawLine(const GSVertexSW* vertex, const uint32* index)
{
const GSVertexSW& v0 = vertex[index[0]];
const GSVertexSW& v1 = vertex[index[1]];
GSVertexSW dv = v1 - v0;
GSVector4 dp = dv.p.abs();
@ -531,7 +540,7 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW2& edge, c
GSVector4 scissor = m_fscissor_x;
top = FindMyNextScanline(top);
while(top < bottom)
{
GSVector8 dy(GSVector4(top) - p0.yyyy());
@ -566,7 +575,7 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW2& edge, c
if(!IsOneOfMyScanlines(top))
{
top += (m_threads - 1) << THREAD_HEIGHT;
top += (m_threads - 1) << m_thread_height;
}
}
@ -721,7 +730,7 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW& edge, co
GSVector4 scissor = m_fscissor_x;
top = FindMyNextScanline(top);
while(top < bottom)
{
GSVector4 dy = GSVector4(top) - p0.yyyy();
@ -758,7 +767,7 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW& edge, co
if(!IsOneOfMyScanlines(top))
{
top += (m_threads - 1) << THREAD_HEIGHT;
top += (m_threads - 1) << m_thread_height;
}
}
@ -810,16 +819,16 @@ void GSRasterizer::DrawSprite(const GSVertexSW* vertex, const uint32* index)
while(top < bottom)
{
r.top = top;
r.bottom = std::min<int>((top + (1 << THREAD_HEIGHT)) & ~((1 << THREAD_HEIGHT) - 1), bottom);
r.bottom = std::min<int>((top + (1 << m_thread_height)) & ~((1 << m_thread_height) - 1), bottom);
m_ds->DrawRect(r, scan);
int pixels = r.width() * r.height();
m_pixels.actual += pixels;
m_pixels.total += pixels;
top = r.bottom + ((m_threads - 1) << THREAD_HEIGHT);
top = r.bottom + ((m_threads - 1) << m_thread_height);
}
}
@ -1132,11 +1141,13 @@ void GSRasterizer::DrawEdge(int pixels, int left, int top, const GSVertexSW& sca
GSRasterizerList::GSRasterizerList(int threads, GSPerfMon* perfmon)
: m_perfmon(perfmon)
{
m_scanline = (uint8*)_aligned_malloc((2048 >> THREAD_HEIGHT) + 16, 64);
m_thread_height = compute_best_thread_height(threads);
m_scanline = (uint8*)_aligned_malloc((2048 >> m_thread_height) + 16, 64);
int row = 0;
while(row < (2048 >> THREAD_HEIGHT))
while(row < (2048 >> m_thread_height))
{
for(int i = 0; i < threads; i++, row++)
{
@ -1161,8 +1172,8 @@ void GSRasterizerList::Queue(const shared_ptr<GSRasterizerData>& data)
ASSERT(r.top >= 0 && r.top < 2048 && r.bottom >= 0 && r.bottom < 2048);
int top = r.top >> THREAD_HEIGHT;
int bottom = std::min<int>((r.bottom + (1 << THREAD_HEIGHT) - 1) >> THREAD_HEIGHT, top + m_workers.size());
int top = r.top >> m_thread_height;
int bottom = std::min<int>((r.bottom + (1 << m_thread_height) - 1) >> m_thread_height, top + m_workers.size());
while(top < bottom)
{
@ -1196,10 +1207,10 @@ bool GSRasterizerList::IsSynced() const
return true;
}
int GSRasterizerList::GetPixels(bool reset)
int GSRasterizerList::GetPixels(bool reset)
{
int pixels = 0;
for(size_t i = 0; i < m_workers.size(); i++)
{
pixels += m_workers[i]->GetPixels(reset);

View File

@ -129,6 +129,7 @@ protected:
IDrawScanline* m_ds;
int m_id;
int m_threads;
int m_thread_height;
uint8* m_scanline;
GSVector4i m_scissor;
GSVector4 m_fscissor_x;
@ -138,7 +139,7 @@ protected:
typedef void (GSRasterizer::*DrawPrimPtr)(const GSVertexSW* v, int count);
template<bool scissor_test>
template<bool scissor_test>
void DrawPoint(const GSVertexSW* vertex, int vertex_count, const uint32* index, int index_count);
void DrawLine(const GSVertexSW* vertex, const uint32* index);
void DrawTriangle(const GSVertexSW* vertex, const uint32* index);
@ -198,6 +199,7 @@ protected:
GSPerfMon* m_perfmon;
vector<GSWorker*> m_workers;
uint8* m_scanline;
int m_thread_height;
GSRasterizerList(int threads, GSPerfMon* perfmon);

View File

@ -274,6 +274,7 @@ GSdxApp::GSdxApp()
m_default_configuration["debug_opengl"] = "0";
m_default_configuration["dump"] = "0";
m_default_configuration["extrathreads"] = "2";
m_default_configuration["extrathreads_height"] = "4";
m_default_configuration["filter"] = "2";
m_default_configuration["force_texture_clear"] = "0";
m_default_configuration["fxaa"] = "0";